JP2020086784A - Manufacturing condition specification system and method - Google Patents

Abstract

To provide a technique capable of specifying a preferable manufacturing condition even when a manufacturing state varies, and sustaining or upgrading product quality.SOLUTION: A computer of a manufacturing condition specification system uses manufacturing condition data and quality data obtained from a manufacturing flow at each of plural times to construct a model concerning each of the manufacturing condition and product quality for each change in a manufacturing state on each of manufacturing processes of the manufacturing flow (S2). The computer uses the models and a quality target value to calculate as first data for each model a predictive value of manufacturing condition data for the next time through learning (S4). The computer uses the models, manufacturing condition data for the current time, and quality data to predict quality data for the next time and calculates a quality error (S5). The computer uses the first data and quality error to specify the manufacturing condition data for the next time through learning (S8).SELECTED DRAWING: Figure 3

Description

The present invention relates to a technology such as information processing, and relates to a technology related to control and specification of manufacturing conditions of a manufacturing flow.

In the manufacturing system in the manufacturing industry, the quality of manufactured products may change according to the setting and control of the manufacturing conditions of the manufacturing flow. Therefore, in order to maintain or improve the quality of the manufactured product, an information processing system or the like (which may be referred to as a manufacturing condition specifying system) for specifying suitable manufacturing conditions has been developed.

Japanese Patent Application Laid-Open No. 2013-84057 (Patent Document 1), Japanese Patent Application Laid-Open No. 2011-39763 (Patent Document 2), and Japanese Patent Laid-Open No. 2013-205890 (Patent Document 3) are examples of prior art relating to the specification of manufacturing conditions. Can be mentioned. Patent Document 1 describes that a probabilistic model is constructed from past manufacturing conditions and a manufacturing condition that matches a target value is calculated in a product quality control method or the like. Patent Document 2 describes that, in an output value prediction method or the like, a plurality of predicted values are output from past performance data. Patent Document 3 describes that, in a machine learning system or the like, a class set expressed in a nonparametric manner is generated when the number of inputs when a model is used is larger than a predetermined number, in other words, the number of dimensions is reduced. There is.

JP, 2013-84057, A JP, 2011-39763, A JP, 2013-205890, A

In the field of manufacturing industry, the manufacturing state is changing every day. The change in the manufacturing state is, for example, a change in the state of the manufacturing apparatus in the manufacturing process. For example, the state of parameters such as current, voltage, temperature, and pressure may change in the manufacturing process by continuing the operation in the manufacturing flow. Further, for example, when maintenance work is performed on the manufacturing apparatus, the state of the manufacturing apparatus may change.

In order to maintain or improve product quality, it is effective to identify suitable manufacturing conditions in response to changes in manufacturing conditions, set the manufacturing conditions in the manufacturing flow, and operate. The product quality is a predetermined evaluation index value, and is obtained, for example, as a value of the inspection result of the quality inspection process, for example, as a yield.

As a manufacturing condition specifying system of the prior art example, there is a system that predicts manufacturing conditions by using a learning model on a computer. This system builds a model using actual data including manufacturing conditions and quality of a manufacturing flow, and predicts suitable manufacturing conditions based on learning by a predetermined learning model.

However, the manufacturing condition specifying system of the prior art example has room for improvement in specifying suitable manufacturing conditions in response to changes in manufacturing conditions. For example, immediately after the change in the manufacturing state, the number of data obtained from the operation results is small, that is, the number of data for advancing learning by the learning model is small. Therefore, it is difficult to improve the prediction accuracy of the model. In order to improve the prediction accuracy of the model, it is necessary to input a certain number of data to the model and proceed with learning, but this takes time. Immediately after the manufacturing state changes, if the prediction accuracy of the model is low, it is not possible to specify a suitable manufacturing condition. As a result, product quality cannot be maintained or improved.

An object of the present invention is to provide a manufacturing condition specifying system capable of specifying a suitable manufacturing condition even when there is a change in the manufacturing state and maintaining or improving product quality. Problems, configurations, effects, and the like other than the above will be described in the modes for carrying out the invention.

A typical embodiment of the present invention has the following configuration. A manufacturing condition specifying system according to an embodiment is a manufacturing condition specifying system including a computer that specifies a manufacturing condition of each manufacturing process of a manufacturing flow, and the computer is a plurality of time points including a current time from the manufacturing flow. A model relating to the manufacturing condition and quality is constructed using manufacturing condition data and quality data, and when the model is constructed, if a change in the manufacturing state of the manufacturing process of the manufacturing flow is included, a plurality of models are created. Each model is constructed for each change in the manufacturing state, and the predicted value of the manufacturing condition data at the next time point is calculated in each model based on the learning in the first learning model using the model and the quality target value. Calculated as one data, using the model and the manufacturing condition data and quality data at the present time, predict the quality data at the next time point, and calculate the quality error between the quality data at the next time point and the quality data at the current time point, Using the first data and the quality error, the manufacturing condition data at the next time point is specified based on the learning by the learning model, and the information including the specified manufacturing condition data at the next time point is stored and output.

According to the representative embodiment of the present invention, it is possible to specify suitable manufacturing conditions and maintain or improve product quality even when there is a change in the manufacturing state in the manufacturing condition specifying system.

It is a figure which shows the structure of the computer of the manufacturing condition identification system of embodiment of this invention. FIG. 3 is a diagram showing a manufacturing flow and a configuration example of a model in the embodiment. It is a figure which shows the process flow of the manufacturing condition specific process in the manufacturing condition specific system of embodiment. It is a figure which shows the structural example of a functional block and data in the manufacturing condition identification system of embodiment. It is a figure which shows the process flow of the subspace conversion process in the manufacturing condition specific|specification system of embodiment. FIG. 6 is a diagram showing an example of a model construction screen in the embodiment. FIG. 6 is a diagram showing an example of a screen of manufacturing condition data and partial space data in the embodiment. It is a figure which shows the example of a screen of quality data in an embodiment. It is a figure which shows the example of a screen of the manufacturing condition data specified in embodiment. FIG. 6 is a diagram showing an example of a model construction setting screen in the embodiment. It is a figure which shows the process example of step S11 in embodiment. It is a figure which shows the process example of step S13 and the process example of step S14 in embodiment. It is a figure which shows the process example of step S14 in embodiment. It is a figure which shows the structural example of manufacturing condition data in embodiment. FIG. 3 is a diagram showing a configuration example of quality data in the embodiment. It is a figure which shows the structural example of causal relationship model data in embodiment. It is a figure which shows the structural example of 1st learning model data in embodiment. FIG. 3 is a diagram showing a configuration example of partial space data in the embodiment. It is a figure which shows the structural example of 3rd learning model data in embodiment. FIG. 6 is a diagram showing an example of model construction from manufacturing condition data according to presence/absence of manufacturing state change in the embodiment.

Hereinafter, embodiments of the present invention will be described in detail with reference to the drawings. In all the drawings for explaining the embodiments, the same parts are designated by the same reference numerals in principle, and the repeated description thereof will be omitted.

(Embodiment)
The manufacturing condition specifying system according to the embodiment of the present invention will be described with reference to FIGS. The manufacturing condition specifying system according to the embodiment is a system that specifies a suitable manufacturing condition corresponding to a change in a manufacturing state of a target manufacturing flow using a model and learning on a computer. The manufacturing condition specifying method according to the embodiment is a method including steps executed in the manufacturing condition specifying system according to the embodiment.

The manufacturing condition specifying system according to the embodiment uses the manufacturing condition data and the quality data at a plurality of time points including the present time, which may include the manufacturing state change of the manufacturing flow, based on the performance of the operation, and relates to the manufacturing condition and the quality. Build one or more models. This system uses the manufacturing condition data and model up to the present time to calculate the predicted value of the manufacturing condition data at the next time based on the learning in the learning model. In addition, this system uses the quality data and the model up to the present time to calculate the predicted value of the quality at the next time and calculates the quality error from the quality at the current time. This system uses those data to identify suitable manufacturing condition data at the next point in time based on the learning by the learning model. As a result, suitable manufacturing condition data for the next time point after the manufacturing state change can be obtained. This system stores and outputs information such as specified manufacturing condition data. The manufacturing condition data is reflected in the manufacturing flow, that is, set, etc., and the operation is performed. As a result, it is possible to maintain or improve the product quality after the manufacturing state has changed.

In addition, the manufacturing condition specifying system according to the embodiment converts the manufacturing condition data obtained by using the model into a subspace so as to reduce the number of dimensions. This conversion is a conversion that reduces the number of dimensions of the data without reducing the information amount of the model and enables subsequent learning. This system uses the subspace data obtained after the conversion to specify a suitable manufacturing condition based on a learning model. At the time of the specific processing, it is only necessary to handle data having a small number of dimensions as an input, and the processing can be efficiently performed.

[Manufacturing condition identification system]
FIG. 1 shows the configuration of a manufacturing condition specifying system according to an embodiment. The manufacturing condition system of the embodiment is implemented by the computer 1. The computer 1 can be configured by a general PC or the like. Alternatively, the computer 1 may be configured by a server device or the like on a communication network. The computer 1 realizes a characteristic function by software program processing. The user operates the computer 1 to perform work such as specification of manufacturing conditions. A user is, for example, a person such as a system engineer (SE) having specialized knowledge about manufacturing systems and machine learning.

The computer 1 has an input/output unit 11, a communication unit 12, a display unit 20, a control unit 30, a storage unit 40, etc., which are connected by a bus or the like not shown. The input/output unit 11 is connected to an input device (for example, a keyboard or a mouse), a display device (for example, a liquid crystal display, a touch panel), or another output device (for example, a printer), which are not shown, and receives a user operation. The communication unit 12 is a portion including a communication interface device for a communication network such as a LAN outside the computer 1, and performs communication processing with an external server device or a device of a manufacturing system. The communication unit 12 acquires data such as manufacturing condition data and information from an external device under the control of the control unit 30.

The display unit 20 configures a screen (corresponding screen data, etc.) related to the manufacturing condition specifying system and displays it on the display screen of the display device. Various screens, which will be described later, display various information such as manufacturing conditions, quality, and models. The screen functions as a graphical user interface (GUI) of the manufacturing condition specifying system. GUI components such as windows, scroll bars, list boxes, and buttons are displayed on the screen, and user operations are possible through them.

The user or the computer 1 acquires necessary data such as manufacturing condition data and quality data from each manufacturing process of the manufacturing flow of the target manufacturing system (FIG. 2 described later). The method of acquiring this necessary data is arbitrary and is not particularly limited. In the embodiment, in particular, the computer 1 is connected to the target manufacturing system via a communication network such as a LAN. Then, the computer 1 acquires the manufacturing condition data, the quality data, the manufacturing flow configuration information, and the like via communication from the manufacturing device, the sensor, the control device, or the like of the manufacturing system. Thereby, the computer 1 can monitor the state of the manufacturing flow and the operation record. The computer 1 can use, for example, a file or a signal output from a manufacturing apparatus or a sensor as data.

The user inputs necessary data to the computer 1 and causes the computer 1 to execute calculation. Calculator 1 specifies suitable manufacturing conditions for the target manufacturing flow by calculation using a model constructed using the input data. The user confirms the suitable manufacturing conditions obtained by the computer 1 on the screen and reflects them on the manufacturing flow of the target manufacturing system. That is, parameter values for control corresponding to the manufacturing conditions are set in the manufacturing apparatus of each manufacturing process of the manufacturing flow. Note that the computer 1 may be configured to be able to transmit and set the manufacturing conditions and the like to the manufacturing apparatus of each manufacturing step of the manufacturing flow via communication.

The control unit 30 is, in other words, a processor, and is configured by known elements such as CPU, RAM, and ROM. The control unit 30 has the following as main processing units implemented based on program processing. That is, the control unit 30 includes a model building unit 31, a model manufacturing condition specifying unit 32, a quality specifying unit 33, a manufacturing condition determining unit 34, a partial space specifying unit 35, and a manufacturing condition specifying unit 36. The control unit 30 stores various data relating to the manufacturing condition specification in the storage unit 40 and manages the information.

The storage unit 40 stores various data and information relating to the manufacturing condition specification. The storage unit 40 can be configured by a non-volatile memory, a storage device, or the like, and may be configured by an external DB server or the like. The storage unit 40 includes a manufacturing condition data storage unit 41, a quality data storage unit 42, a model storage unit 43, a first learning model storage unit 44, a second learning model storage unit 45, a partial space storage unit 46, and a third learning model storage. Including part 47. The manufacturing condition data storage unit 41 stores the manufacturing condition data acquired from the manufacturing flow and the manufacturing condition data specified by the computer 1. The quality data storage unit 42 stores quality data and the like acquired from the manufacturing flow. The model storage unit 43 stores model data constructed by the computer 1 and related to manufacturing conditions and quality. The first learning model storage unit 44 stores data of a first learning model described later. The second learning model storage unit 45 stores data of a second learning model described later. The partial space storage unit 46 stores partial space data described later. The third learning model storage unit 47 stores data of a third learning model described later.

The display unit 20 is a unit that performs a process of displaying various data and information on the screen of the display device based on the process of the control unit 30. The display unit 20 includes a model display unit 21, a manufacturing condition display unit 22, a partial space display unit 23, a quality display unit 24, and a manufacturing condition specifying display unit 25. The model display unit 21 graphically displays the constructed model on a screen (FIG. 6 described later). The manufacturing condition display unit 22 displays the manufacturing condition data on the screen (FIG. 7 described later). The subspace display unit 23 displays the subspace data on the screen (FIG. 7 described later). The quality display unit 24 displays a graph of product yield as quality on the screen (FIG. 8 described later). The manufacturing condition specification display unit 25 displays the suitable manufacturing condition data specified by the computer 1 on the screen (FIG. 9 described later).

The model building unit 31 uses the manufacturing condition data in the manufacturing condition storage unit 41 and the quality data in the quality data storage unit 42 to perform a process of building one or more models (model 50 in FIG. 4 described later). This is a part (step S2 in FIGS. 3 and 4 described later). The model construction includes updating of the already constructed model.

The model manufacturing condition specifying unit 32 uses the model obtained by the model constructing unit 31 and the quality target value, and based on learning in a predetermined learning model (referred to as a first learning model), the manufacturing condition at the next time. Is a part for performing the processing for specifying the predicted value of (step S4 in FIGS. 3 and 4 described later).

The manufacturing condition determination unit 34 uses the latest manufacturing condition data obtained from the actual manufacturing without using the above model, and based on learning in a predetermined learning model (referred to as a second learning model). , Which is a part for performing a process of determining the predicted value of the manufacturing condition at the next time (step S6 in FIGS. 3 and 4 described later).

The quality specifying unit 34 predicts the quality at the next time using the model obtained by the model building unit 31 and the manufacturing condition data at the current time, and a quality that is an error between the quality at the next time and the quality at the current time. This is a part for performing a process of calculating an error (step S5 in FIGS. 3 and 4 described later).

The subspace specifying unit 35 receives the next-time manufacturing condition data (referred to as first data) obtained from the model manufacturing-condition specifying unit 32 and the next-time manufacturing condition data (second data) obtained from the manufacturing-condition determining unit 34. Yes) and the quality error obtained from the quality specifying unit 33 are used as input data. The partial space specifying unit 35 is a part that performs a process of converting the manufacturing condition data of the input data into a partial section that is a space different from the original space (step S7 in FIGS. 3 and 4 described later). In other words, the subspace conversion process, which is this process, is a process of projecting input data onto a low-dimensional space with a reduced number of dimensions. This subspace conversion process reduces the number of dimensions of the input manufacturing condition data so as to match the input format of the manufacturing condition specifying process of the manufacturing condition specifying unit 36. This subspace conversion processing reduces the number of input dimensions to match the number of input dimensions of the third learning model described below. As a result of this partial space conversion processing, partial space manufacturing condition data that is partial space data is obtained.

The manufacturing condition specifying unit 36 receives the partial space manufacturing condition data obtained from the partial space specifying unit 35 as an input, and finally based on learning in a predetermined learning model (referred to as a third learning model), finally This is a part for performing the process of specifying the optimum manufacturing condition (step S8 in FIGS. 3 and 4 described later).

[Manufacturing flow and model]
FIG. 2 shows a manufacturing flow and a configuration example of a model (model 50 in FIG. 4). The upper part of FIG. 2 shows an outline of the manufacturing flow, and the lower part shows an outline of a causal relationship model as an example of a model constructed based on the manufacturing flow. The manufacturing flow is composed of a plurality of processes (=manufacturing processes) from upstream to downstream, and in this example, has four processes #1, #2, #3, #4=#L. In this example, the last process #L is a quality inspection process. One or more manufacturing apparatuses and one or more sensors are provided in association with each process. An inspection device and a sensor are provided in the quality inspection process, and the quality data is output as inspection result data. For example, the process #1 has manufacturing apparatuses #1 and #2. The manufacturing apparatus #1 controls and executes the manufacturing in the process #1 according to the set manufacturing conditions. The manufacturing apparatus #1 includes sensors A and B, for example. The sensor A detects a predetermined parameter value at the time of manufacturing by the manufacturing apparatus #1 in the process #1, and outputs it as observation data. Similarly, a manufacturing apparatus and a sensor are provided in each process, and they are connected according to the order of processes and the like. The structure of such a manufacturing flow is managed as manufacturing flow configuration information by the manufacturing system.

Manufacturing condition data is associated with each process. The manufacturing condition data (corresponding data item) is a parameter value that controls the states of the manufacturing apparatus and the sensor, and examples of general parameters include current, voltage, temperature, and pressure. The computer 1 of the manufacturing condition specifying system according to the embodiment acquires manufacturing condition data, quality data, manufacturing flow configuration information, and the like from the manufacturing flow of the manufacturing system. The user or the computer 1 can acquire the manufacturing condition data and the like from the manufacturing device or the control device of each process. The manufacturing condition specifying system can acquire the manufacturing condition data and the quality data as a monitor of the operation results at each time point in time series including the manufacturing state change of the manufacturing flow.

In addition, in the embodiment, a method in which a quality inspection step is provided as the last step of the manufacturing flow, quality data is obtained from the quality inspection step, and the quality data and the manufacturing condition data are associated with each other is applied. The present invention is not limited to this, and can be similarly applied to, for example, a method in which a quality inspection flow exists independently of the manufacturing flow.

The manufacturing condition specifying system according to the embodiment builds a causal relationship model based on the manufacturing condition data and the quality data acquired from the manufacturing flow. The causal relationship model on the lower side of FIG. 2 can be expressed by a network structure. That is, this model can be represented by the connection between the node indicated by the circle and the edge indicated by the arrow. Each manufacturing condition data can be expressed as a node. The edge represents the direction of the causal relationship. In this example, the case of a causal relationship model using observation data of the sensors A, B,..., R as the manufacturing condition data is shown. For example, in the part corresponding to the process #1, the node A corresponding to the observation data of the sensor A is connected to the node B and the node C. Node B is connected to node C. The node C is connected to the nodes D and G.

The manufacturing condition specifying system according to the embodiment specifies suitable manufacturing condition data based on the constructed model, and stores and outputs the specified manufacturing condition data. It should be noted that the user or the computer 1 refers to the manufacturing flow configuration information and the like to confirm the correspondence relationship such as which manufacturing process and manufacturing device of the manufacturing flow the specified manufacturing condition data and the like are associated with. it can. For example, the user can output information in which the manufacturing condition data specified by the computer 1 is associated with the structure of the manufacturing flow to a person in the manufacturing system or the manufacturing site. Alternatively, the computer 1 can transmit the information in which the specified manufacturing condition data is associated with the structure of the manufacturing flow to the manufacturing system and set it in each manufacturing apparatus or the like.

[Process flow and functional block configuration]
FIG. 3 shows a processing flow including manufacturing condition specifying processing by the control unit 30 of the computer 1 in the manufacturing condition specifying system of the embodiment. The processing flow of FIG. 3 includes steps S1 to S9. FIG. 4 shows the configuration of each data and the functional block of the control unit 30 in association with each step of FIG. The processing will be described below in the order of steps with reference to FIG. 4 together with FIG.

(S1) First, in step S1, the control unit 30 acquires manufacturing condition data from each manufacturing process of the manufacturing flow and acquires quality data from the quality inspection process as a monitor of operation results. The control unit 30 stores the acquired manufacturing condition data in the manufacturing condition data storage unit 41, and stores the acquired quality data in the quality data storage unit 42. The model construction unit 31 refers to the acquired manufacturing condition data D1 and quality data D2 from the storage unit 40.

In addition, each data (D1, D2) is data acquired at each time point (corresponding acquisition time) on the time series including the time points before and after the manufacturing state change. Time T is used as a point in time for explanation. The manufacturing condition data and the quality data obtained at a certain time T may be expressed as the manufacturing condition data at the time T and the quality data at the time T. In addition, the current time may be represented as time (t) and the next time may be represented as time (t+1).

(S2) Next, in step S2, the model construction unit 31 obtains at the time of the past operation stored in the storage unit 40 in addition to the manufacturing condition data and quality data at the latest time T obtained in step S1. The model 50 is constructed using the obtained manufacturing condition data and quality data. The model 50 is one or more models. In the embodiment, the model 50 is a plurality (N) of models when there is a manufacturing state change. Let N be the number of models. In the embodiment, a causal relationship model as shown in FIG. 2 is used for all of the plurality (N) of models.

The model building unit 31 builds each model of the model 50 from the manufacturing condition data D1 and the quality data D2 at each time according to the manufacturing state change. The model building unit 31 builds a plurality of models by dividing each model at each time when the manufacturing state changes (for example, FIG. 20 described later). At the time of model building, the model building unit 31 divides the manufacturing condition data according to the event time when there is a manufacturing state change event in a certain manufacturing process. Then, the model building unit 31 builds a model for each section according to the sections of the plurality of divided data. For example, when there is one event of manufacturing state change, each model of the model before the manufacturing state change and the model after the manufacturing state change is constructed.

It should be noted that the model 50 is not limited to the causal relationship model, and models of other methods can also be applied. Further, a plurality of (N) models may include a mixture of models of a plurality of types.

The model construction unit 31 stores the data of the constructed model 50 in the model storage unit 43. The model display unit 21 also displays the information of the constructed model 50 on the model construction screen of FIG.

(S3) In step S3, the control unit 30 sets a quality target value based on a user operation. For example, the display unit 20 provides a quality setting screen. A quality target value setting field may be provided in the quality screen of FIG. 8 described later. The user sets a quality target value, for example, a product yield target value on the setting screen. The control unit 30 causes the quality data storage unit 42 to store the setting information including the quality target value D3. If there is preset quality target value setting information, the control unit 30 may refer to it.

(S4) In step S4, the model manufacturing condition identifying unit 32 uses the model 50 constructed in S2, and if there are a plurality of models, sets the model to the next time (t+1) with respect to the current time (t) for each model. Identify predicted values for manufacturing condition data. In FIG. 4, the model manufacturing condition specifying unit 32 calculates the manufacturing condition data D6 at the next time as a predicted value for each model from the model 50 and the quality target value D3 based on learning in the first learning model LM1. .. The manufacturing condition data D6 is the first data.

At the time of the processing of S4, the model manufacturing condition specifying unit 32 specifies, for each model of the model 50, the manufacturing condition that the quality is good with respect to the quality target value D3. The model manufacturing condition identifying unit 32 stores the obtained manufacturing condition data D6 in the storage unit 40. The model manufacturing condition specifying unit 32 stores the data of the first learning model used, including the update, in the first learning model storage unit 45. In the embodiment, for example, a reinforcement learning model is used as the first learning model.

(S5) In step S5, the quality specifying unit 33 predicts the quality data D7 at the next time (t+1) from the manufacturing condition data D4 and the quality data D5 at the current time (t) using the model 50 in S2. Then, the quality specifying unit 33 compares the predicted quality data D7 at the next time with the quality data D6 at the current time (t), and calculates a quality error D8 which is an error between them. As the manufacturing condition data D4 and the quality data D5 at the current time (t), the data stored in the quality data storage unit 42 can be used. The quality specifying unit 33 stores the data (D7, D8) calculated in S5 in the quality data storage unit 42.

(S6) In step S6, the manufacturing condition determination unit 34 uses the model 50, and based on the learning in the second learning model LM2 from the manufacturing condition data D4 at the latest current time (t) in the actual result, The predicted value of the manufacturing condition data D9 at time (t+1) is determined. This manufacturing condition data D9 is the second data. The manufacturing condition determination unit 34 stores the obtained manufacturing condition data D9 in the storage unit 40. The manufacturing condition determination unit 34 stores the data of the second learning model LM2 including the update in the second learning model storage unit 45. In the embodiment, as the second learning model LM2, for example, a reinforcement learning model is used similarly to the first learning model LM1.

In the embodiment, both the manufacturing condition data D6 which is the first data and the manufacturing condition data D9 which is the second data are used as the input data of the process of S7.

(S7) In step S7, the subspace identifying unit 36 generates the manufacturing condition data D6 at the next time, which is the first data obtained at S4, and the manufacturing condition data D9 at the next time, which is the second data obtained at S6. Each data including the quality error D8 obtained in S5 and the manufacturing condition data D4 at the current time is input. The subspace specifying unit 36 performs a subspace conversion process for converting the manufacturing condition data of the input data including the first data and the second data into the subspace data. The next-time manufacturing condition data D6, which is the first data obtained in S4, is N pieces of data corresponding to the number N of the models 50. The manufacturing condition data D9 at the next time, which is the second data obtained in S6, is one piece of data. That is, the manufacturing condition data input in S7 is (N+1) pieces of data. The subspace conversion process of S7 is a process of projecting the (N+1) pieces of manufacturing condition data onto a subspace (in other words, a low-dimensional space) that is a space different from the original space.

The conversion of S7 is a conversion that reduces the number of dimensions of the input data so as to match the number of dimensions of the input format of the manufacturing condition specifying processing of S8. In other words, the number of dimensions of the input format of the third learning model LM3 in the manufacturing condition specifying processing of S8 is matched with the number of dimensions of the subspace data obtained in S7. The subspace data obtained in S7 is data projected on the subspace, and is data in which the number of dimensions of the parameter is reduced. When the number of dimensions of the input (N+1) pieces of manufacturing condition data is DN1 and the number of dimensions of the subspace data is DN2, DN1>DN2. The number of dimensions DN2 is matched with the number of dimensions of the input of the third learning model LM3 in the manufacturing condition specifying process of S8. Details of the process of S7 will be described later.

The partial space specifying unit 36 obtains the partial space manufacturing condition data D10 as the partial space data which is the output data. The subspace identification unit 36 stores the obtained subspace data in the subspace storage unit 46. The subspace display unit 23 displays the subspace data in the subspace data area 230 of the screen of FIG. 7.

(S8) In step S8, the manufacturing condition specifying unit 36 uses the partial space manufacturing condition data D10 obtained in S7, and based on learning by the third learning model LM3, the manufacturing condition at the optimum next time (t+1). The data D11 is specified. This manufacturing condition data D11 is the third data. During this process, the manufacturing condition specifying unit 36 finally specifies the optimum manufacturing condition data to be applied to the manufacturing flow from the (N+1) manufacturing condition data in the partial space data. In this process, the manufacturing condition specifying unit 36 constructs the third learning model LM3 using the past manufacturing condition data and the quality error D8. In the embodiment, a reinforcement learning model is used as the third learning model LM3. The manufacturing condition specifying unit 36 causes the manufacturing condition data storage unit 41 to store the specified manufacturing condition data D11 at the next time (t+1). The manufacturing condition identifying unit 36 stores the data of the third learning model LM3 including the update in the third learning model storage unit 47.

(S9) In step S9, the display unit 20 updates the screen display content so that various data and information such as manufacturing condition data, quality data, and model obtained as a result of the above processing are displayed on the screen. To do. For example, the manufacturing condition specification display unit 25 displays the information of the manufacturing condition data D11 at the next time (t+1) specified in S8 in the manufacturing condition data area 251 of the result screen of FIG. In addition, the manufacturing condition specification display unit 25 displays the information of the model selected from the model 50 corresponding to the specified manufacturing condition data D11 in the model area 252 of the result screen of FIG. The user can check the optimum manufacturing conditions and the model used to specify the optimum manufacturing conditions by looking at the result screen.

The user can reflect the specified optimum manufacturing condition data D11 in the manufacturing flow of the manufacturing system according to the operation. For example, by pressing the OK button in the result screen of FIG. 9, the manufacturing condition data D11 can be set in the manufacturing flow. After that, the operation after the next time is performed. The processing flow of FIG. 3 is similarly repeated at each time point, and accordingly, learning in each learning model progresses, and the prediction accuracy gradually increases. The screen display of various data can be omitted according to a user operation or user setting. The user can specify desired data on the screen and display it on the screen, or can output the data to the outside.

[Subspace conversion processing]
FIG. 5 shows a process flow of the subspace conversion process by the subspace identifying unit 35 in step S8 of the process flow of FIG. The processing flow of FIG. 5 includes steps S10 to S15.

(S10) First, in step S10, the subspace identifying unit 35 acquires the manufacturing condition data D4 at the current time (t) in FIG. Further, the subspace identifying unit 35 acquires the first data, which is the N pieces of manufacturing condition data D6 at the next time (t+1) obtained from the model 50 in S4. Further, the subspace identifying unit 35 acquires the second data, which is the one piece of manufacturing condition data D9 at the next time (t+1) obtained from the latest manufacturing condition data D4 in S6. Further, the subspace identifying unit 35 acquires the quality error D8 obtained in S5.

(S11) In step S11, the subspace identifying unit 35 calculates the difference between the manufacturing condition data of the current time (t) and the next time (t+1) from the manufacturing condition data (D6, D9) obtained in S10, The difference is held as a vector value (in other words, a difference vector). Details of S11 are shown in FIG. 11 described later.

(S12) In step S12, the subspace identifying unit 35 concatenates the vector values obtained in S11 into vertical vectors, as shown in FIG.

(S13) In step S13, the subspace identifying unit 35 expresses the vertical vector obtained in S12 by the sum of the vectors for each item of the manufacturing condition data, as shown in FIG.

(S14) In step S14, the subspace identifying unit 35 calculates the coefficient value of each vector so that each component of the decomposed vector becomes 1 in the vector sum obtained in S13. Details of S14 are shown in FIG. The subspace specifying unit 35 regards the coefficient value of the obtained vector as data obtained by projecting the manufacturing condition data on the subspace, and sets it as the partial space manufacturing condition data D10. This subspace data is used as an input for the process of S8.

FIG. 11 shows a processing example of step S11. In this example, briefly, a case where the number of data items of the manufacturing condition data is two and the number N of the models 50 is 2 is shown. Definitions are as follows. Each model of the model 50 is represented by a model Mi (i=1, 2). In this example, model M1=model #1 and model M2=model #2 are used. Two times T corresponding to before and after the manufacturing state change are set to time (t) and time (t+1). The manufacturing condition data at the time (t) of the model Mi is x _i,t . The manufacturing condition data at time (t+1) of the model Mi is x _i,t+1 . It is assumed that i=0 (x ₀ ) represents the manufacturing condition data specified from the actual manufacturing condition data. The time difference (that is, the difference vector) for each model of the manufacturing condition data is Δx. The calculation for an example of one data value is as follows.

For the model #1 (M1), the manufacturing condition data x _{1, t+1 at} time (t+1) is expressed by Equation 1 when expressed by a matrix of 2 rows and 1 column, in other words, a column vector with the number of dimensions=2. .. Expression 1 is x _{1, t+1} =(4,4) in the row vector representation, and the first data item value=4 and the second data item value=4. The manufacturing condition data x _{1,t at} time (t) is expressed by Equation 2 when expressed similarly. Expression 2 is x _1,t 2 =(2,3) in the row vector expression, and the first data item value=2 and the second data item value=3. The subspace specifying unit 35 takes the difference between the manufacturing condition data x _1,t+1 of Expression ₁ and the manufacturing condition data x _{1,t of} Expression 2, and creates a difference vector of Expression 3. Expression 3 is Δx ₁ =x _{1, t+1} −x _1,t =(2,1) in the row vector expression.

Similarly, for the model #2 (M2), the manufacturing condition data x _{2,t+1 at} time (t+1) is expressed by Equation 4 when expressed by a matrix of 2 rows and 1 column, in other words, a column vector with the number of dimensions=2. expressed. Expression 4 is x _2,t+1 =(5,4) in the row vector expression. The manufacturing condition data x _{2,t at} time (t) is expressed by Equation 5. Expression 5 is x _2,t =(3,3) in the row vector expression. Subspace identification unit 35 obtains a difference between the manufacturing condition data x _{2, t} manufacturing condition data x _{2, t + 1} and Formula 5 Formula 4, to create a difference vector equation 6. Expression 6 is Δx ₂ =x _2,t+1 −x _2,t =(2,1) in the row vector expression.

Regarding the actual manufacturing condition (i=0), the manufacturing condition data x _{0,t+1 at} the time (t+1) is expressed by Equation 7 when expressed by a matrix of 2 rows and 1 column, in other words, a column vector of the number of dimensions=2. expressed. Expression 7 is x _0,t+1 =(5,4) in the row vector expression. The manufacturing condition data x _{0,t at} time (t) is represented by Expression 8. Expression 8 is x _0,t =(3,3) in the row vector expression. Subspace identification unit 35 obtains a difference between the manufacturing condition data x _{0, t} manufacturing condition data x _{0, t + 1} and Formula 8 in Formula 7, to create a difference vector equation 9. Expression 9 is Δx ₀ =x _0,t+1 −x _0,t =(2,1) in the row vector expression.

FIG. 12A shows a processing example of step S12. The subspace specifying unit 35 creates a vertical vector in which the difference vectors {Δx ₁ , Δx ₂ , Δx ₀ } of FIG. 11 are vertically connected. Let V be the vertical vector. The vertical vector V is expressed by Expression 10. The vertical vector V in Expression 10 is (2, 1, 2, 1, 2, 1) in the row vector expression.

FIG. 12B shows a processing example of step S13. The subspace specifying unit 35 expresses the vertical vector V of FIG. 12A by the sum of the vectors for each data item of the manufacturing condition data. Let V1 be the vector of the first data item and V2 be the vector of the second data item. V=V1+V2, which is represented by Expression 11. In the row vector expression, V1 in Expression 11 is (2,0,2,0,2,0), and V2 is (0,1,0,1,0,1).

FIG. 13 shows a processing example of step S14. The subspace specifying unit 35 calculates coefficient values so that each component becomes 1 for the vector sum (V=V1+V2) in FIG. The coefficients are c1 and c2. The calculation formula is represented by Formula 12. Expression 12 is V=V1+V2=c1×(1,0,1,0,1,0)+c2×(0,1,0,1,0,1) in the row vector expression. The coefficient value obtained from Expression 12 is, for example, (c1, c2)=(2, 1). As shown in Expression 13, the subspace identifying unit 35 sets the obtained coefficient value as a subspace data in a matrix of 2 rows and 1 column, in other words, a column vector of the number of dimensions=2. The subspace data of Expression 13 is (a1, a2) in the row vector expression. This subspace data is data obtained by projecting manufacturing condition data so as to reduce the number of dimensions. The number of dimensions of the original vector V=6, whereas the number of dimensions of the subspace data=2.

The manufacturing condition specifying unit 36 in step S8 of FIG. 4 specifies the manufacturing condition data D11 at the next time on the basis of the third learning model LM3 by inputting the partial space manufacturing condition data D10 having such a reduced number of dimensions. Perform processing to Since the number of dimensions of the input data is reduced by the subspace conversion, the processing of S8 can be efficiently performed in a short time as compared with the case where the subspace conversion is not performed. That is, the manufacturing condition specifying system according to the embodiment can specify the optimum partial condition in a short time. Since the optimum partial condition can be immediately reflected in the manufacturing flow, the product yield can be improved in a short period even after the manufacturing state is changed.

[Change in manufacturing status]
The following are examples of manufacturing state changes. Maintenance work may be performed by a worker on a manufacturing apparatus or the like of each manufacturing process that constitutes the manufacturing flow of the manufacturing system. In this case, before or after the maintenance event (corresponding event time), the internal or external physical state of the manufacturing apparatus, for example, the value of a parameter such as current may change instead of a constant value. is there. There is a possibility that the optimum manufacturing conditions have changed according to this change in the manufacturing state. When the manufacturing conditions are changed and the same manufacturing conditions as those before the change are continuously applied and the operation is continued, the quality of the product manufactured after the change may be deteriorated. That is, there is a possibility that the suitable manufacturing conditions may have changed internally according to changes in the manufacturing state, and it is necessary to specify the changed suitable manufacturing conditions. Therefore, the manufacturing condition identification system of the embodiment has a function of constructing a model and a learning model in response to such a manufacturing state change and specifying a suitable or optimal manufacturing condition corresponding to the next time after the change. Have

Note that the unit of time and time handled by the manufacturing condition specifying system of the embodiment is a unit having a size corresponding to the product manufacturing time of the target manufacturing system, such as day (day), hour (hour), and minute. A unit such as (minute) can be given and can be set appropriately.

The parameters forming the manufacturing condition data are in accordance with the manufacturing flow of the target manufacturing system, can be set appropriately, and are not particularly limited. Examples of parameters are current, voltage, temperature, pressure and the like. These parameters are parameters that can be controlled or measured. For example, the current and voltage at a predetermined location in the manufacturing apparatus, the pressure in a predetermined space in the manufacturing apparatus, the temperature of the environment inside or outside the manufacturing apparatus, and the like can be mentioned. Examples of applicable manufacturing systems and manufacturing flows include, but are not limited to, at least semiconductor manufacturing systems and manufacturing flows.

[Display screen]
6 to 10 show examples of main screens displayed on the display screen of the display device by the display unit 20 of the computer 1 in the manufacturing condition specifying system. Examples of this screen include the model construction screen of FIG. 6, the manufacturing condition screen of FIG. 7, the quality screen of FIG. 8, the result screen of FIG. 9, the model construction setting screen of FIG.

[Model construction screen]
The model construction screen of FIG. 6 includes one or more model display areas 210 for displaying the information of the model 50. In the example of FIG. 6, three model display areas 210 corresponding to three of the plurality (N) of models are shown. The model display area 210 is an area in which the model configuration obtained by the model construction unit 31 is graphically displayed by the model display unit 21. The model display area 210 includes an area 213, a setting button 211, a build button 212, and a label 214. One model is displayed in the area 213. The label 214 displays a number, a name, etc. as a label for each model.

In the example of FIG. 6, the network structure of the causal relationship model is displayed in the area 213 as an example of the model. In the area 213, the causal relationship model is represented by a plurality of nodes and edges. Each node corresponds to manufacturing condition data. A node ID (corresponding manufacturing condition data ID) or the like is displayed on each node. If the entire model cannot be displayed on the screen, a part of the model can be displayed using a scroll bar or the like, and the display part can be changed by a user operation. For example, model #1 is composed of 305 pieces of data represented by ID=X1 to X304 and Y as a plurality of nodes. The last data of ID=Y represents quality data.

[Model building setting screen]
When the user presses the setting button 211 on the model construction screen of FIG. 6, the model construction setting screen of FIG. 10 is displayed by pop-up or the like. The model construction setting screen of FIG. 10 shows an example of a screen for the user to input and set information related to the construction of the model 50. This screen example is a screen example when the causal relationship model is used, but when a model of another system is used, a setting screen corresponding to that system is provided. This screen has a setting column 261 for use data and a setting column 262 for causal relationship model construction conditions. The setting column 261 can set manufacturing condition data or the like, which is data (for example, data monitored from the manufacturing flow) used when the causal relationship model is constructed, by a method of referring to a file or the like. The setting column 262 can be set by using a method of referring to a setting file of the condition, a method of setting while confirming on a viewer (another setting screen), or the like when constructing the causal relationship model. .. In the case of the method set on the Viewer, the discretization method, the structure learning algorithm, the use/non-use of constraint conditions, etc., which are generally used in the case of a causal relationship model, can be set by a method of selecting from a selection branch. Known methods can be used for these methods.

The user presses the OK button after setting the conditions and the like. As a result, the control unit 30 reflects the conditions and the like in the manufacturing condition specifying system. Then, the causal relationship model is constructed under the conditions and the like by the processing of the model constructing unit 31. When the user presses the Cancel button, the conditions and the like are not reflected, and the state (screen in FIG. 6) before the setting screen is opened is returned to.

The model display unit 21 constitutes one or more model display areas 210 for displaying one or more causal relationship models constructed by the model construction unit 31 and displays them in the model construction screen of FIG. 6. When a plurality of models are constructed, the model display unit 21 gives a label 214 to each model and displays it. The label 214 is associated with the model ID of the data in the storage unit 40.

[Production condition screen]
The manufacturing condition screen of FIG. 7 includes a manufacturing condition display area 220 and a partial space display area 230. The manufacturing condition display area 220 is an area for displaying the acquired manufacturing condition data stored in the manufacturing condition data storage unit 41. The manufacturing condition display area 220 is, for example, in a table format, and includes a time and a plurality of manufacturing condition data as a column of items. The “time” item corresponds to the time T, which is the acquisition time, and values are stored in chronological order. For example, the latest acquisition time is time Tn. The manufacturing condition data item 221 displays a plurality of data items of the respective manufacturing condition data. Here, the number of manufacturing condition data is m, and a case where m=304 is shown, for example.

The partial space display area 230 is an area for displaying the partial space data obtained by the partial space specifying section 35 by the partial space display section 23. In the partial space display area 230, the time difference (that is, the difference vector) of the manufacturing condition data for each model is displayed as the partial space data. In this example, the case of the partial space data regarding the manufacturing condition data X1 is shown. The subspace display area 230 is, for example, in a table format, and includes a model ID and a plurality of times as a column of items. The model ID is an ID for each model and corresponds to the label. The plurality of times are the times between the times T. For example, regarding the manufacturing condition data X1 and the model #1, the difference between the data item values is 2 from the time T1 to the time T2, and the difference between the data item values is 0 from the time T2 to the time T3. is there.

[Quality screen]
The quality screen of FIG. 8 includes a product yield display area 240. The product yield display area 240 is an area in which the data of a graph in which the quality data stored in the quality data storage unit 42 is converted into a product yield and arranged in chronological order is displayed. The product yield display unit 24 constitutes this product yield graph and displays it in the product yield display area 240.

Further, in the product yield display area 240, it is also possible to display the event time 242 which is data representing the time of the event that occurred in the manufacturing process of the manufacturing flow. This event includes an event such as maintenance corresponding to a manufacturing state change. For example, the event time Tx indicates the time at which a maintenance event in a certain manufacturing apparatus Dx has occurred.

As shown on the quality screen, the product quality fluctuates over time. In this example, the yield tended to decrease around a certain event time 242 (Tx). After that, when suitable manufacturing conditions were specified by the manufacturing condition specifying system of the embodiment and applied to the manufacturing flow, the yield tended to increase and was improved.

[Result screen]
The result screen of FIG. 9 includes a manufacturing condition specifying display area 250. The manufacturing condition specifying display area 250 is an area for displaying the optimum manufacturing condition data at the next time specified by the manufacturing condition specifying unit 36, and the model corresponding thereto. The manufacturing condition specification display unit 25 displays the optimum manufacturing condition data at the next time in the area 251, for example, in a table format. Further, the manufacturing condition identification display unit 25 displays the model selected from the model 50 corresponding to the optimum manufacturing condition data at the next time in the area 252 in the same format as the model display area 210 of FIG. 6, for example. ..

[Manufacturing condition data]
FIG. 14 shows a table of a configuration example of the manufacturing condition data stored in the manufacturing condition data storage unit 41. The table of FIG. 14 has a product ID, an acquisition time, and a plurality of parameters 1401 as columns of items corresponding to the header information. The plurality of parameters 1401 are a plurality of data items that constitute the manufacturing conditions, and have P1 “temperature 1”, P2 “temperature 2”, P3 “pressure 1”, and P4 “pressure 2” as examples of parameters. The “product ID” is an identifier for each product manufactured in the manufacturing process. The “acquisition time” is the time when the computer 1 acquired the manufacturing condition data.

It should be noted that in the table of manufacturing condition data as shown in FIG. 14, the information on the event time of the manufacturing state change may be stored, or the information for identifying the manufacturing state change may be stored for each data item of the manufacturing condition data. It may be configured to.

[Quality data]
FIG. 15 shows a table of a configuration example of the quality data stored in the quality data storage unit 42. The table of FIG. 15 has a product ID, a quality value, and a quality inspection result as a column of items. The “quality value” is a value representing the quality of the product obtained in the quality inspection step (FIG. 2), and one example is the product yield. The “quality inspection result” is a value obtained as a result of a predetermined determination based on the “quality value” in the quality inspection process. The determination is, for example, a comparison determination with a threshold value. For example, when the quality value satisfies a target value (for example, a threshold value of 0.90 or more), the value of the quality inspection result is “good”, and when the quality value does not satisfy it, it is “bad”.

[Model data]
FIG. 16 shows a configuration example of data of the model 50 stored in the model storage unit 43. The case of a causal relationship model is shown as an example of the model. As shown on the upper side of FIG. 16, the causal relationship model can be represented in a tabular form. In the rows and columns of the table, parameters (“temperature 1” and the like in FIG. 14 and quality values of quality data) corresponding to each node of the manufacturing condition data are arranged. In the cell portion where the matrix intersects, a value representing a causal relationship (connection relationship by edges) between the node in the row and the node in the column is stored. This value is binary, and the value is “0” when the nodes are not connected (that is, there is no causal relationship) and the value is “0” when the nodes are connected (that is, there is a causal relationship and there is an edge). 1” is set.

The lower side of FIG. 16 shows a case where the upper side table is represented as a network structure. The model storage unit 43 stores the data in the upper table. When displayed on the screen, the network structure can be configured by conversion from the table. In addition, the model data shown in FIG. 16 exists for each model construction of the manufacturing state change in step S2, and a label or ID for identification is given to each model and saved as a part of the model data.

[First learning model]
FIG. 17 shows a configuration example of data of the first learning model stored in the first learning model storage unit 44. In this example, a deep learning model is used as an example of a reinforcement learning model that is a first learning model. The first learning model data can be represented in a tabular format as shown on the upper side. This table includes a left side table 1701 and a right side table 1702. The table 1701 on the left side has a layer number and the number of nodes in a layer as a column of items. "Layer number" is a number that identifies a layer. “Number of nodes in layer” indicates the number of nodes in the layer. The table 1702 on the right side has layer numbers, weight numbers, and weights as columns of items. The “weight number” is a number for identifying the “weight” held in each layer. “Weight” indicates the numerical value of the weight given to the edge between the nodes. The data of such a table is stored in the first learning model storage unit 44. The lower side of FIG. 17 shows a case where the upper side table is represented in the network structure format. The number N of layers shown in FIG. 17 is different from the number N of the above model. The configuration of the second learning model data is the same as that of the first learning model data.

[Subspace data]
FIG. 18 shows a table of a configuration example of the partial space data stored in the partial space storage unit 45. A table 1801 on the upper side of FIG. 18 is a table representing the difference vector (step S11 of FIG. 5, for example, FIG. 11) calculated in the processing of the subspace identifying unit 36. This table 1801 has a model number and a difference for each data item (parameter such as “temperature 1”) of manufacturing condition data as a column of items. The model number is a number corresponding to a label or ID for identifying each of the plurality (N) of models 50. The difference vector for each mode is stored for each row of the table. This table 1801 has the number of dimensions and the amount of data according to the product of the number N of models and the number of data items.

A table 1802 on the lower side of FIG. 18 shows a table of a configuration example of data of coefficient values (step S14 of FIG. 5, for example, FIG. 13) calculated in the process of the subspace identifying unit 36. This data is a set of coefficient values for each manufacturing condition data item and corresponds to subspace data. This table 1802 has a coefficient value for each data item (parameter such as “temperature 1”) of manufacturing condition data as a column of items. This table 1802 has a smaller number of dimensions and a smaller amount of data than the table 1801.

As described above, the subspace specifying unit 36 processes the data of the difference vector shown in the upper table 1801, calculates the coefficient values shown in the lower table 1802, and calculates the subspace data that is the calculated coefficient value, It is stored in the subspace storage unit 46.

[Third learning model]
FIG. 19 shows a configuration example of the data of the third learning model stored in the third learning model storage unit 47. Similarly, as an example of the reinforcement learning model which is the third learning model, a case where a deep learning model is used is shown. The configuration example of the third learning model data can be represented by a table or a network structure, like the configuration example of the first learning model data of FIG. Note that the number of layers and the like differ for each learning model. The upper table of FIG. 19 includes a left table 1901 and a right table 1902. The data in the upper table is stored in the third learning model storage unit 47.

[Example of model construction according to changes in manufacturing status]
FIG. 20 shows, as a supplement, an outline and an example when the model 50 is constructed from the manufacturing condition data according to the manufacturing state change. FIG. 20A shows the construction of a single model when there is no event of manufacturing state change. Table 2001 shows an example of manufacturing condition data. As in the table 2001, for example, there is a row for each “acquisition time” as the manufacturing condition data regarding the product of “Product ID”=A001. As the latest time point, for example, the data of “acquisition time”=“10/1/10:50” (10:15 on October 1) is acquired. At this time, a single causal relationship model is constructed by using the data at the latest time point and the data at a plurality of past time points (for example, a preset number such as 1000). The “acquisition time” is the time when the computer 1 acquires the manufacturing condition data, and correspondingly, the manufacturing flow has information such as the time when the manufacturing in the manufacturing process is executed, A form using such information may be used.

FIG. 20B shows the construction of a plurality of models when there is a manufacturing state change event. The data contents of table 2001 are the same. The manufacturing flow of the manufacturing system also has data on the time of the event in which the manufacturing state change occurred. The computer 1 can also acquire the data of the event time from the manufacturing system. It is assumed that the time of the event of the manufacturing state change is, for example, the time "10/1/10:15" in the event E1 and "10/1/10:35" in the event E2. An event E1 at time "10:15" occurs between the time (10:10) on the second line and the time (10:20) on the third line. An event E2 at time "10:35" occurs between the time (10:30) on the fourth line and the time (10:40) on the fifth line. The model construction unit 31 divides the data at the event time and constructs a model for each divided period. For example, the model #1 is constructed using a plurality of data including the data in the first and second rows. Model #2 is constructed using a plurality of data including the data in the third and fourth rows. Model #3 is constructed using a plurality of data including the data in the fifth and sixth rows. Note that the number of rows of manufacturing condition data for constructing one model is not limited to two, and is possible.

[Learning model]
The learning methods that can be applied to the above learning models are as follows.

First learning model LM1: state space model, reinforcement learning model Second learning model LM2: state space model, reinforcement learning model Third learning model LM3: reinforcement learning model.

Known reinforcement learning is a type of machine learning that deals with a problem in which an agent in a certain environment observes a current state and decides an action to be taken next. Reinforcement learning involves learning strategies that maximize the reward through a series of actions. A known deep learning model or the like can be applied as the reinforcement learning model. The known state space model is a type of time series analysis model. The state space model is composed of a state model and an observation model.

[Effects, etc.]
As described above, according to the manufacturing condition specifying system of the embodiment, it is possible to specify suitable manufacturing conditions even when there is a change in the manufacturing state, and maintain or improve product quality. In particular, according to the embodiment, it is possible to identify the optimum manufacturing condition corresponding to the quality target value in response to the manufacturing state change. According to the embodiment, it is possible to specify suitable manufacturing conditions in a short period of time immediately after the manufacturing state changes and even when the number of operation data is small. Further, in particular, according to the embodiment, by using the subspace conversion, even when there are many manufacturing state changes, even when the manufacturing conditions and the number of parameters of the model are large, it is easy to handle, and suitable manufacturing conditions can be set. Can be specified. According to the embodiment or the modified example, according to the priority policy, it is possible to improve the quality at an early stage after the manufacturing state is changed, or to improve the prediction accuracy of the model and then improve the quality.

As described above, the manufacturing condition specifying system according to the embodiment builds a plurality of models from the past manufacturing condition change state and the manufacturing condition data acquired at each time point. This system estimates the manufacturing condition at the next time from each of the plurality of models. This system takes a difference for each time between time points from a plurality of estimated manufacturing conditions and embeds it in a subspace based on the difference. This system uses the subspace data, which is the data embedded in the subspace, as input, and specifies the optimum manufacturing condition at the next time based on learning.

The following is also possible as another embodiment. First, as the manufacturing condition specifying system of the modified example, a mode in which the subspace conversion process of step S7 of FIG. 3 is omitted is possible. In this case, in the manufacturing condition specifying process of step S8, each data of the results of steps S4, S5 and S6 is input. This modification is also effective when the manufacturing state changes little or when the manufacturing conditions and the number of model parameters are relatively small.

As the manufacturing condition specifying system of the modified example, it is possible to use only the first data obtained as a result of step S4 as the manufacturing condition data to be the input data in the subspace conversion process of step S7 of FIG. In the above-described embodiment, both the first data resulting from step S4 and the second data resulting from step S6 are used in step S7. The second data resulting from step S6 is the manufacturing condition data D9 predicted using the manufacturing condition data D4 at the latest time point (t). In this case, the manufacturing condition data D11 at the next point in time will be specified by emphasizing the latest manufacturing data. Therefore, this form is effective in the case of a policy that emphasizes early improvement of product quality immediately after a change in the manufacturing state.

On the other hand, in the modified example, the second data as a result of step S6 is not used during the processing of steps S7 and S8. This form is effective in the case of a policy that emphasizes improving the prediction accuracy of the model rather than increasing the product quality immediately after the change in the manufacturing state at an early stage. In this modification, it is possible to improve the prediction accuracy of the model by advancing learning for a certain period of time after the manufacturing state has changed, and then to specify the optimum manufacturing condition data with high accuracy.

[Comparative example]
The differences between the manufacturing condition specifying system of the embodiment and the system of the prior art as a comparative example are as follows. Patent Document 1 describes that a probabilistic model is constructed from past manufacturing conditions in order to specify manufacturing conditions that are non-defective products, and manufacturing conditions that match target values are calculated. However, with this technology, when a manufacturing state change that cannot be expressed by past manufacturing conditions occurs, it is necessary to modify the constructed model.

Patent Document 2 describes that a plurality of prediction values are output from past performance data (for example, manufacturing conditions and qualities), and the output values are weighted according to the degree of similarity to perform prediction. However, this technique cannot specify the manufacturing condition when the prediction target (here, quality) is a good product.

Patent Document 3 describes that, using reinforcement learning, a condition (for example, a manufacturing condition) that satisfies a certain condition (for example, a condition for a product to be a good product) is searched for in an environment (for example, a manufacturing process). However, in this technique, when there are a plurality of models to be used, the number of items of manufacturing conditions to be specified is reduced by class classification. Therefore, with this technique, it may be possible that the manufacturing conditions for obtaining a good product cannot be obtained.

On the other hand, the manufacturing condition specifying system according to the embodiment uses the manufacturing conditions and qualities obtained from the past manufacturing including the manufacturing state change to build a plurality of models, and the manufacturing conditions and quality errors obtained from each model are calculated. It can be used to search for optimal manufacturing conditions.

Although the present invention has been specifically described above based on the embodiments, the present invention is not limited to the above-described embodiments and can be variously modified without departing from the scope of the invention.

DESCRIPTION OF SYMBOLS 1... Calculator, 30... Control part, 31... Model construction part, 32... Model manufacturing condition specific part, 33... Quality specific part, 34... Manufacturing condition determination part, 35... Partial space specific part, 36... Manufacturing condition specific part.

Claims (15)

A manufacturing condition specifying system including a computer for specifying the manufacturing condition of each manufacturing process of a manufacturing flow,
The calculator is
Using the manufacturing condition data and quality data at a plurality of time points including the present time from the manufacturing flow, build a model relating to the manufacturing condition and quality,
In the case of including a change in the manufacturing state of the manufacturing process of the manufacturing flow when constructing the model, each model is constructed for each change in the manufacturing state as a plurality of models,
Using the model and the quality target value, the predicted value of the manufacturing condition data at the next time point is calculated as the first data in each model based on the learning in the first learning model,
Using the model and the current manufacturing condition data and quality data, predict the next time quality data, calculate the quality error between the next time quality data and the current quality data,
Using the first data and the quality error, based on the learning in the learning model, to identify the manufacturing condition data at the next time point,
Storing and outputting information including the specified next time manufacturing condition data,
Manufacturing condition identification system. In the manufacturing condition specifying system according to claim 1,
The calculator is
Using the manufacturing condition data at the present time point, the predicted value of the manufacturing condition data at the next time point is calculated as the second data based on the learning by the second learning model,
Using the first data, the second data, and the quality error, the manufacturing condition data at the next time point is specified based on learning by the learning model,
Manufacturing condition identification system. In the manufacturing condition specifying system according to claim 1,
The calculator is
Converting the first data into subspace data to reduce the number of dimensions,
Using the subspace data and the quality error, based on learning in the learning model, to specify the manufacturing condition data at the next time point,
Manufacturing condition identification system. In the manufacturing condition specifying system according to claim 1,
The calculator divides the manufacturing condition data at the plurality of time points into data of a plurality of periods according to the time of occurrence of a predetermined event for each of the manufacturing steps as the change of the manufacturing state, and the data of the divided periods. Build each model for each
Manufacturing condition identification system. In the manufacturing condition specifying system according to claim 4,
The event includes maintenance work for the manufacturing apparatus of each manufacturing process,
Manufacturing condition identification system. In the manufacturing condition specifying system according to claim 1,
The model includes a causal model,
Manufacturing condition identification system. In the manufacturing condition specifying system according to claim 1,
The first learning model includes a state space model or a reinforcement learning model,
Manufacturing condition identification system. In the manufacturing condition specifying system according to claim 2,
The second learning model includes a state space model or a reinforcement learning model,
Manufacturing condition identification system. In the manufacturing condition specifying system according to claim 1,
The learning model includes a reinforcement learning model,
Manufacturing condition identification system. In the manufacturing condition specifying system according to claim 1,
The computer displays each model of the model on a display screen of a display device,
Manufacturing condition identification system. In the manufacturing condition specifying system according to claim 1,
The computer displays the manufacturing condition data at the plurality of time points and the partial space data on a display screen of a display device,
Manufacturing condition identification system. In the manufacturing condition specifying system according to claim 1,
The computer causes the display screen of the display device to display quality data at the plurality of time points, and the time point of occurrence of a predetermined event for each manufacturing process as the manufacturing state change,
Manufacturing condition identification system. In the manufacturing condition specifying system according to claim 1,
The computer displays, on a display screen of a display device, a model selected from the specified next time manufacturing condition data and the respective models of the model associated with the specified next time manufacturing condition data. Let
Manufacturing condition identification system. In the manufacturing condition specifying system according to claim 1,
The calculator is
Obtaining manufacturing condition data and quality data at a plurality of time points including the present time from the manufacturing flow,
The specified next-time manufacturing condition data is transmitted to a manufacturing system that controls the manufacturing flow and set in each of the manufacturing steps.
Manufacturing condition identification system. A manufacturing condition specifying method in a manufacturing condition specifying system including a computer for specifying the manufacturing condition of each manufacturing step of the manufacturing flow,
As the steps executed in the computer,
Using the manufacturing condition data and quality data at a plurality of time points including the present time from the manufacturing flow, a model relating to the manufacturing condition and quality is constructed, and when the model is constructed, the manufacturing state of the manufacturing process of the manufacturing flow. In the case of including a change of, each step of building each model as a plurality of models for each change of the manufacturing state,
Calculating the predicted value of the manufacturing condition data at the next time point as the first data in each model based on the learning in the first learning model using the model and the quality target value;
Using the model and the current manufacturing condition data and quality data, predicting quality data at the next time point, and calculating a quality error between the quality data at the next time point and the quality data at the current time point,
Using the first data and the quality error to specify manufacturing condition data at the next time point based on learning by a learning model,
Storing and outputting information including the specified next time manufacturing condition data,
And a manufacturing condition specifying method.

Images